# Evidence, with its limits

Reviewed September 24, 2026. This page separates release proof, named tests, maintainer runs and research specifications. [Machine-readable claims](claims.json) and [schema](claims.schema.json) retain exact source/test bindings and falsifiers. A passing reference validator is not a certification of semantic truth.

## Released baseline

Camden Work v0.1.0: commit `f8ffad19a198915564b9816e156b22b3fd4dddee`. Documentation baseline: `b48a377b9675dc64a03b0957fe5e56d4c8fac1e8`. [Public release CI](https://github.com/prohormonePro/camden-work/actions/runs/35888011009) and the [September 23 addendum](verification-2026-09-23.md) retain their original scope. The 84 release-suite tests, 18 overlapping installed checks and seven challenge predicates have different denominators.

The seven challenge predicates are `crash_after_target_commit`, `one_original_effect`, `unknown_observed`, `retry_fenced`, `independent_task_verified`, `registered_wait_serviced`, and `unsafe_baseline_duplicates`. The two child exits are 72. One original effect is not the total target row count. The naive baseline uses new identities and is not an industry benchmark.

## Source and negative tests

`tests/test_finalization.py` rejects export without adoption, adoption without delivery, altered inbox bytes, enlarged work scope and lost current visibility. `tests/test_continuity.py` exercises conditional compensation and fresh verification. These narrow mappings are PUBLIC_SOURCE_AND_TEST_MAPPED, not evidence that every larger gauntlet row ran together.

## New observed local run

The separately identified source-tree benchmark has four started Windows episodes: two cold-controller and two recovery-of-recovery. All four reached BUSINESS_COMPLETE within their declared local boundary. Exact exits were `[72,0]` and `[72,73,0]`. Both predeclared obligations had one effect; original budgets/deadlines and attempts survived; old owner use was rejected; local adoption, delivery bytes and finalization passed.

This is MAINTAINER_REPORTED with [raw manifest and every episode](https://github.com/prohormonePro/camden-work/tree/main/benchmarks/runs/2026-09-24-windows). The plan binds actual source SHA-256, runtime, storage, repetitions and fault schedule. No unsuccessful started episode was omitted. The runner is [reproducible from source](continuity-gauntlet.md); the immutable released wheel is unchanged.

B_e and D_e were zero within those synthetic runs. A_e was two or three harness launches per episode. Setup human time was NOT_MEASURED. Automatic restart, background wake, remote effects and genuinely new business decisions were not exercised. No autonomous-supervisor claim follows from these observations.

## What could disprove the claims

Find a duplicate original effect, a stale owner accepted at the effect boundary, lost or reset attempt/deadline state, independent eligible work blocked, or a local READY with missing adoption/delivery/current evidence under the stated trust model. A correctly configured alternative meeting the same contract defeats an exclusivity claim. [Related work](related-work.md) has source review only; no comparative run exists.

The catalog has 283 research mechanisms. The [versioned differentiation overlay](differentiation.json) maps six narrow relationships; all other mappings remain UNKNOWN. A mapping is neither a whole-mechanism certification nor a count of prevented failures.